A Deep Learning Based Surrogate Model for Thermal Signature Prediction in Laser Metal Deposition

Laser metal deposition (LMD) is an additive manufacturing method for metal parts by using focused thermal energy to fuse materials as they are deposited. During LMD, transient thermal signatures such as the in-situ thermal images of melt pool, contain rich information about process performance. Early prediction of such transient thermal signatures provides opportunities for process monitoring and defect prevention. While physics-based models of LMD have been conventionally used for thermal signature prediction, they have limitations and are computationally expensive for real-time prediction. A scalable, efficient data-science-based model is therefore needed. This paper develops a deep-learning-based surrogate model, called LMD-cGAN, to predict and emulate the transient thermal signatures in LMD. The model generates images for the thermal dynamics of melt pool conditionally on the deposition layer. It enables early prediction of future-layer thermal signatures for an in-process part based on its early-layer thermal signatures. To respect the physics in LMD, a physics-guided image selection (PGIS) mechanism is integrated with LMD-cGAN to calibrate the predictions against physical benchmarks of transient melt pool for the process. The effectiveness and efficiency of the proposed method are demonstrated in a case study on the LMD of Ti-4Al-6V thin-walled structures. Note to Practitioners—With online sensing, many LMD applications have real-time process data that convey valuable information about the process status and part quality. The proposed method leverages these data for thermal signature prediction. LMD-cGAN is a deep-learning-based surrogate model that learns the population profile of real thermal signatures and generates thermal signatures from there. The proposed PGIS mechanism in LMD-cGAN ensures the physical validity of these predictions by benchmarking them against physical insights about the process. LMD-cGAN can be applied to predict thermal signatures in future layers based on early-layer thermal signatures of an in-process part (an implicit assumption here is that the in-process part to be predicted for is the same type). LMD-cGAN can also be applied to emulate thermal signatures in specific layers. To generate thermal signatures for generic, non-defect parts, the training data should be selected with caution – the part where the data were collected should have no obvious defects, so the thermal signatures generated by LMD-cGAN show the regular thermal dynamics. Compared with pure physical models, the proposed method incorporates process uncertainties captured from the early-layer data, hence “on-the-fly” emulation of the melt pool, while characterizing the inherent relationship between the LMD process and thermal signatures.